[LUCENE-10302] PriorityQueue: optimize where we collect then iterate by using O(N) heapify

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Lucene Fields: New
• Assignee: Unassigned
• Reporter: David Smiley
• Created: 09/Dec/21 13:23
• Updated: 28/Aug/22 16:32

Description

Looking at LUCENE-8875 (LargeNumHitsTopDocsCollector.java ) I got to wondering if there was faster-than O(N*log(N)) way of loading a PriorityQueue when we provide a bulk array to initialize the heap/PriorityQueue. It turns out there is: the JDK's PriorityQueue supports this in its constructors, referring to "This classic algorithm due to Floyd (1964) is known to be O(size)" – heapify() method. There's another that may or may not be the same; I didn't look too closely yet. I see a number of uses of Lucene's PriorityQueue that first collects values and only after collecting want to do something with the results (typical / unsurprising). This lends itself to a builder pattern that can look similar to LargeNumHitsTopDocsCollector in terms of first having an array used like a list and then move over to the PriorityQueue if/when it gets full (it may not).

Issue Links

• relates to: LUCENE-8875 Should TopScoreDocCollector Always Populate Sentinel Values?